AI for FP&A means using forecasting models, assistants and AI agents in financial planning and analysis: pulling and reconciling data, forecasting, explaining variances and drafting the monthly pack. It can take much of the assembly off the month. The FP&A lead still owns the assumptions, the reading of each variance and the number that goes to the board.
Finance has bought a great deal of planning software, and the average budget still takes 8.7 weeks to produce. The Association for Financial Professionals reported that figure from its 2026 FP&A Benchmarking Survey of 332 finance practitioners, published on 20 January 2026, and found it unchanged from three years earlier. Only 43% of the organisations it surveyed ran rolling forecasts.
The month has the same shape at a smaller scale. The ledger closes, the actuals land, and the first days go on pulling, reconciling and formatting before anybody asks why a number moved.
Planning vendors sell AI in FP&A as one thing, and it is three: a model that forecasts, an assistant that answers questions, and an agent team that produces finished work. Each fits a different part of the month, and each fails in a different way.
In this article
- What can AI do for FP&A?
- How is AI used in FP&A, and how widely?
- What is the difference between a forecasting model, an assistant and an AI agent?
- Which FP&A work suits an AI agent, and which does not?
- Do AI agents belong in analytical work, or only in transactional work?
- How do you check the figures in an AI-prepared finance pack?
- How does AI change the month for an FP&A team?
- How should FP&A treat travel and expense spend?
- Frequently asked questions
- What would a team of specialist agents take off your month?
What can AI do for FP&A?
AI can take four jobs in financial planning and analysis: preparing data, producing a baseline forecast, explaining what moved, and drafting the commentary and pack. The first two are established uses of forecasting models and automation. The last two are where AI agents come in, and they are only safe with a checkpoint and a citation behind every figure.
In practice, the four jobs look like this:
- Preparing data: pulling actuals from the ledger, card and expense feeds, mapping them to cost centres and reconciling the totals.
- Forecasting: projecting the coming periods from history and drivers, with a range around the central figure.
- Explaining: splitting each movement against budget into volume, price, mix and timing, and flagging those that cross a threshold.
- Drafting: writing the variance commentary and assembling the monthly pack.
A product sold as an "AI financial analyst", or as AI for CFOs, is doing one or more of these four jobs, so ask which.
How is AI used in FP&A, and how widely?
Most FP&A practitioners do not use AI regularly, and finance as a whole uses it more for the close and transactions than for planning. In the Association for Financial Professionals' 2025 benchmarking survey, 23% of FP&A practitioners used AI daily, weekly or monthly. In Gartner's 2025 survey, forecasting and scenario planning trailed payables automation and anomaly detection.
Gartner asks a different question: whether a finance function uses AI at all. Gartner's 2025 AI in Finance Survey of 183 CFOs and senior finance leaders, published on 18 November 2025, found 59% reporting AI in their finance function, against 58% in 2024 and 37% in 2023.
The table below takes nine of the twelve use-case figures Gartner published with that survey, from a base of 152 finance leaders, and sorts each by the side of finance it serves.
The three planning-side uses all sit below 30%, and the three most common serve the close or both sides at once. Finance has put AI to work first where its output is easiest to check.
Inside FP&A, AFP found that AI use "has not reached a level of commonality". The finding comes from AFP's 2025 FP&A Benchmarking Survey Report on technology and data, drawn from 362 FP&A practitioners worldwide and published in January 2025, in which another 40% were testing AI with plans to implement it within a year. Because Gartner asked whether a function uses AI at all and AFP asked how often practitioners use it, the two figures do not compare directly.
What is the difference between a forecasting model, an assistant and an AI agent?
A forecasting model returns a number, an assistant returns an answer to the question you typed, and an AI agent team returns a finished piece of work, such as a monthly variance pack, planned and assembled across several steps. They fit different parts of the FP&A month, so ask any vendor which one it is selling.
A forecasting model is statistics applied to your history and drivers, blind to anything that history has never seen, such as an acquisition or a change of travel policy. An AI agent team plans and assembles the work itself, and it is only as trustworthy as the trail behind its figures.
The line between an agent team and the automation finance already runs is drawn in how agentic workflows differ from automation.
The table below maps recurring FP&A work to the kind of AI that fits it, and to the part that stays with a person whichever tool does the rest.
The expensive mistake is asking an assistant for a monthly pack: what comes back looks like one, and nobody planned it or checked it.
Which FP&A work suits an AI agent, and which does not?
FP&A work suits an AI agent when it recurs in the same shape, draws on data traceable to a system of record, and ends in a draft a person can review faster than they could rebuild it. Explaining monthly variances and drafting the pack usually pass. Setting targets, allocating budget and presenting to the board do not.
The Review-Not-Redo Test
The Review-Not-Redo Test is PredictX's five-question check for deciding whether an FP&A deliverable can be handed to an AI agent. Its premise is in the name: delegation pays only when checking the draft costs less than rebuilding it, and when the check would catch a wrong figure. Answer each question yes or no for one deliverable, such as the monthly cost-centre pack.
Without the written threshold in question three, a fluent draft explains every movement with the same confidence, and the one that matters reads like the rest.
The monthly variance commentary for a cost-centre pack usually scores five. The annual allocation of budget between business units usually scores one: it recurs, but it is a negotiation, and nobody can check a negotiation by reading a draft of it.
Do AI agents belong in analytical work, or only in transactional work?
AI agents belong in analytical work when three controls are in place: a plan a person approves before anything runs, a checkpoint where surprising findings stop for review, and a citation behind every figure. Without those controls, keeping agents to transactional work is the right call, because an analytical error arrives as plausible prose and slips past the reader.
The appetite for agentic AI in finance is plain. In Deloitte's Q4 2025 CFO Signals survey of 200 North American CFOs at companies with at least US$1 billion in revenue, published on 13 January 2026, 87% expected AI to be extremely or very important to their finance department's operations in 2026. And 54% said integrating AI agents in their finance departments would be a "transformation priority".
The case for keeping agents out of analysis is a serious one all the same. A transactional flow has a right answer: an invoice matches its purchase order or it does not. A variance can have three plausible causes, and an agent that picks one fluently is more dangerous than one that fails loudly.
Where the case goes wrong is in treating an analytical deliverable as analysis all the way through. Most of a monthly variance pack is assembly: pulling actuals, mapping cost centres, splitting each movement into volume, price and mix, charting and drafting. That assembly has right answers, and the judgement sits in a handful of calls about which movement matters, what caused it and what to do.
Each control keeps the agents on the assembly:
- The plan comes first. Nothing runs until the FP&A lead has approved the method, which is the most basic form of keeping a human in the loop.
- Surprises stop. Anything above the materiality threshold pauses for a person before any commentary is written.
- Every figure cites its source, so a reviewer can check it instead of taking it on trust.
An agent that cannot show where a figure came from has no place in an FP&A pack, however good its prose.
How do you check the figures in an AI-prepared finance pack?
Check three things in an AI-prepared pack: that every figure cites the query and the system it came from, that the pack's totals tie back to the ledger for the same period, and that the method matches last month's or declares what changed. Then trace a small sample of figures to source yourself before the pack goes out.
Fluent output is the hazard: a wrong figure inside a well-written paragraph reads like a right one.
Provenance, for a finance pack, is the recorded trail from each figure back to the query, system and period that produced it, and it turns checking from rework into reading. Orchestra, PredictX's AI agent team for travel and expense, ties every figure to its source query through a full citation trail, the third of its three validation tiers. Whatever tool prepares the pack, ask for that trail by name, then run a short trace of your own.
A short trace before the pack goes out
- Tie the headline totals to the ledger or trial balance for the same period and entity.
- Pick five figures at random, including one quoted in the commentary, and follow each citation to its source query.
- Re-run one of those queries yourself and compare the result.
- Confirm that the period, currency and cost-centre mapping match last month's pack, or that the change is declared.
- Read the commentary for any cause stated as fact that no figure supports, and mark it for the owner of that line.
If step two fails once, the pack waits. That rule does more for trust than any accuracy figure on a product page.
How does AI change the month for an FP&A team?
AI moves the first days after close from building the pack to reviewing it. The pack can re-run on a schedule or when new actuals land, a reforecast becomes cheap enough to run more often, and the team's time shifts towards the conversations with budget holders that no model can hold for them.
Recurring deliverables change first, because a monthly pack keeps the same structure every time: the method is agreed once and each run applies it to new data. Orchestra, for example, can re-run a project weekly, monthly, quarterly or on an event trigger when new data lands, and for FP&A the event trigger matters most, because the pack starts when the actuals arrive.
The reforecast changes next. A rolling forecast means rebuilding the outlook every month or quarter, which is analyst time most organisations in the AFP survey above do not spend. When a model re-runs the baseline whenever a driver moves, that cost falls towards the time it takes to review the result.
Some of the month does not move at all. The close belongs to accounting, and the budget round is a negotiation between people with competing claims on the same money. PredictX's view is that the budget cycle AFP measured shortens only as far as the assembly inside it, because agreement runs at the speed of the people agreeing.
How should FP&A treat travel and expense spend?
Treat travel and expense as a driver-based line, forecast from trips and unit costs rather than rolled forward from last year. It moves with headcount, events and fares, and bookings made ahead of travel show next month's spend before the ledger does. Explain its variances by volume, price and mix, after separating timing differences from overspend.
Travel is hard to forecast because the market underneath it moves. In its outlook of 21 July 2025, the Global Business Travel Association cut its 2025 growth projection for global business travel spending from 10.4%, forecast a year earlier, to 6.6%, citing trade policy uncertainty.
If the industry's own forecast can move almost four points in a year, a travel budget rolled forward from last year's actuals carries assumptions nobody has written down.
Travel also runs on four clocks. A trip is booked, taken, settled on a card and claimed on expenses, and those four events can fall in different months. The ledger sees the last of them, so part of any month-end travel variance is timing.
Separate the timing first, then use the booking data as a leading indicator: air and hotel booked for next month are spend the ledger has not recorded yet. What remains splits three ways:
- Volume: the change in the number of trips or nights.
- Price: the change in the average cost of each, holding the mix constant.
- Mix: the shift between routes, cabins, regions and categories.
The three lead to different conversations. Volume is a question for the budget holder, price for the market, and mix for travel policy. Negotiated airline and hotel rates feed the price assumption as an input; negotiating them is procurement's work.
One of Orchestra's example projects is budget variance analysis, run every month: period-accurate by cost centre and decomposed into volume, price and mix, so the FP&A lead reads the result.
Frequently asked questions
Will AI replace FP&A teams?
AI takes over tasks inside FP&A and leaves the team in place. Pulling, reconciling and first drafts can move to AI; the decisions stay, meaning which variance matters, what caused it, what the business should do and how to explain it to the board. A team that hands over the assembly gains time for those decisions.
What is agentic AI for FP&A?
Agentic AI for FP&A is AI that takes a finance goal, such as explaining the month's variances, drafts a plan, splits the work across specialist agents and returns a finished deliverable for a person to review. It differs from a forecasting model, which returns a number, and from an assistant, which answers one question at a time.
What should you look for in AI for FP&A?
In AI for FP&A, look for three things: a citation from every figure to the query and system behind it, a checkpoint where surprising findings stop for a person before anything is written, and a method that stays the same each month unless someone approves a change. Without those, the output is fluent and uncheckable.
How do you use AI in FP&A?
To use AI in FP&A, start with one recurring deliverable, such as the monthly cost-centre pack. Agree the method and a written materiality threshold, let the AI assemble the draft from traceable data, and review it against the ledger before it goes out. Add a second deliverable once the first has run cleanly for a few cycles.
What are the main AI use cases in FP&A?
The main AI use cases in FP&A are data preparation, baseline and rolling forecasts, scenario comparison, variance explanation and drafting the monthly pack. Data preparation and forecasting suit automation and forecasting models. Variance explanation and drafting suit AI agents, provided every figure carries a citation and surprising findings stop for a person before any commentary is written.
Does AI for FP&A need clean data first?
AI for FP&A needs traceable data more than perfect data. In the Association for Financial Professionals' 2025 benchmarking survey, 61% of FP&A practitioners named unreliable data and 60% named inaccessible data as what held their technology back. An AI that cites its sources exposes bad data quickly, which is uncomfortable and useful in equal measure.
What should a CFO ask before letting AI prepare the monthly pack?
Before handing the monthly pack to AI, ask four questions: can every figure be traced to its source query, do the pack's totals tie to the ledger for the period, who sees a surprising finding first, and who signs the pack before it leaves finance? If any answer is vague, wait.
What would a team of specialist agents take off your month?
A team of specialist agents can take the assembly: pulling and reconciling the actuals, splitting each variance into its drivers, and drafting a cited pack for the FP&A lead to review. The decisions stay with people when the team is built so that no agent decides alone, which is the standard to hold any product to.
Orchestra is PredictX's AI agent team for travel and expense: an AI project manager plans the work, directs nine specialist agents and returns a finished, cited deliverable. Nothing runs until a person approves the plan. Its FP&A Agent covers budgets, forecasts and spend visibility, and the project manager agent dispatches whichever other specialists a project needs.
Its agents never decide alone: surprising findings pause for your review, and nothing commits spend without a named human. Orchestra launched in June 2026 and is deploying across enterprise travel programmes, which is a short record to judge by, so judge the mechanism instead: the first project runs on your own programme data during the walkthrough.
Related Posts
Human in the loop: where should you sit when AI agents do the work?
Agentic workflows, and how they differ from the automation you run
AI agents for procurement: what they can run, and what the team keeps
